2015-07-08: Autonomy Incubator PI Profiles: Loc Tran

Dr. Loc Tran, a recent graduate of Old Dominion University with a PhD in machine learning, is one of the Autonomy Incubator's (AI) newer hires—he's been here since January. In that time, however, he's quickly made himself indispensable as he leads the AI's obstacle avoidance ("tree-dodging") research and serves as the machine learning guru for all the activities that go on here in Building 1222.

If you've followed the AI's demonstrations on Twitter or here on the blog, you'll recognize Loc's work instantly: he's the reason we have an entire artificial forest set up on one half of the flight range. Basically, his work involves using a machine learning algorithm to teach UAVs how to detect and avoid stationary obstacles as they navigate in GPS-denied environments, and since dodging tree trunks under the forest canopy is a compelling example of an environment with no mapping and lots of obstacles, "tree-dodging" came to be the hallmark of Loc's work at the AI.

"I started with about five trees," he said. "Then we put out an ad for old Christmas trees on [the NASA Langley Research Center website] and now..." He gestured toward the small forest filling the far side of the room from his desk. There are about fifteen trees there now, with more arriving every week. Big ones, small ones, spruces, palms; it's a polymer arboretum.

This isn't even all of them; there are some recent deliveries in the back room.
Thanks NASA Langley!

With the capabilities Loc is developing, UAVs could become vital partners to humans in a wide scope of applications, from collecting samples in remote forests for scientists to performing sweeps of wooded areas in search-and-rescue missions. If you're interested in the method of how Loc uses his forest to train the machine learning algorithm on the UAVs, we already asked him about it in a previous post—how convenient is that? Just click here.

Loc explains tree-dodging to Deputy Administrator Dava Newman
and Center Director Dave Bowles during last month's demo

While his research has been both scientifically exciting and entertaining so far, he's looking forward to replacing the safe-but-limited COTS (Commercial Off The Shelf) vehicles he's been using with the Green Machines. With the new, custom-built UAVs' entirely onboard computing capabilities, he says, he'll be able to stage tests and get results that simply are not possible with COTS (Commercial Off The Shelf) vehicles this small.

"I'm anticipating our new vehicle because the problems I'm experiencing with the AR. Drones will... be solved," he said.

While he and the rest of the AI wait eagerly for the Green Machines to be declared research-ready, Loc has turned his attention to a specific facet of obstacle avoidance: computer vision. In contrast to intern Alex Hagiopol's research in visual odometry, which has focused on SVO and PTAM methods, Loc has high hopes for an inertial computer vision algorithm he's been investigating called MSCKF (Multi-State Constraint Kalman Filter).

Rather than relying solely on visual information, MSCKF uses both a camera and an IMU (Inertial Measurement Unit) to collect data about position, acceleration, and rotational velocity. The sensors in smartphones that rotate the display when the phone tilts sideways are IMUs, for example—they sense acceleration and changes in direction. Then, MSCKF employs an algorithm called an extended Kalman filter to filter out the unnecessary "noise" from those two sources and synthesize all the relevant information into an idea of where an object is in space.

When talking about algorithms and concepts that are this complex, it's easy to get tangled up in the jargon and lose sight of just how really, really amazing MSCKF could be. Fortunately, many people doing this kind of research are putting videos of their tests online—the below clip came from Dr. Mingyang Li in California.



"It's a very accurate way of saying, 'This is where you've been and this is where you are now,'" Loc said.  If implemented, then MSCKF would give the Autonomy Incubator's UAVs "unlimited range" and "total independence from the ViCon."

Between teaching UAVs to navigate autonomously and looking for new ways to expand the Autonomy Incubator's research into real-world applications, Dr. Loc Tran's first months at the AI are clearly the beginning of a promising research trajectory.